Abstract

INTRODUCTION: Breast cancer is a commonly diagnosed malignancy among women globally and represents a major public health burden. Breast cancer screening mammography detects nearly 35% of ductal carcinoma in situ (DCIS) among asymptomatic women. Early detection helps with timely treatment and reduces the burden of this disease. Artificial intelligence (AI) is a potential tool to improve screening efficiency. It is being integrated into breast cancer screening to address workforce constraints, while maintaining diagnostic performance. Studies evaluating AI performance for DCIS detection within screening workflows remain limited. OBJECTIVE: To evaluate the diagnostic efficacy, workload impact, and recall rates of AI-integrated screening workflows compared with traditional human double-reading approaches in detecting DCIS in breast cancer screening programmes. METHODS: A meta-analysis was performed. We searched for studies evaluating AI-integrated screening workflows vs. standard double reading in asymptomatic breast cancer screening populations. The primary outcome was the DCIS detection rate per 1,000 women screened (or per 1,000 screenings). Secondary outcomes included recall rate and workload reduction. Risk of bias was assessed using QUADAS-AI. RESULTS: Five studies were included in the analysis. There was no significant difference in DCIS detection between AI-assisted workflows and standard double reading (pooled RR 1.17, 95% CI 0.79-1.75, I² = 82.3%, p = 0.433), with substantial heterogeneity. The outcomes for recall rates were not significantly different overall with AI use (pooled RR 1.19, 95% CI 0.87-1.62, I² = 98.6%, p = 0.275). There was a mean workload reduction of 42.9% (SD 1.5%) with AI integration. CONCLUSION: AI-integrated screening workflows can reduce workload and result in slight increases in recall rates (I² = 98.6%). However, the effect on DCIS detection (I² = 82.3%) is not clear due to heterogeneity across studies. There is a need to exercise caution in incorporating AI into the diagnostic process for DCIS until more data becomes available.

This work is licensed under a Creative Commons Attribution 4.0 License.